TMCnet News

Surface Quality Prediction and Processing Parameters Determination on Atmospheric Pressure Plasma Arc Cleaning [Sensors & Transducers (Canada)]
[April 22, 2014]

Surface Quality Prediction and Processing Parameters Determination on Atmospheric Pressure Plasma Arc Cleaning [Sensors & Transducers (Canada)]


(Sensors & Transducers (Canada) Via Acquire Media NewsEdge) Abstract: The theory of Least Squares Support Vector Machines was applied to metal surfaces cleaning by atmospheric pressure plasma arc. An intelligent predictive model of the non-linear relationship between cleaning quality and process parameters was established with the k-fold cross training of sample data. An orthogonal experiment was conducted to assess the effect of processing parameters on surface quality. The experimental results and predicted values show that the atmospheric pressure plasma arc (APPA) cleaning is effective in reducing considerably the amount of lubricant. Furthermore, it is feasible to apply LS-SVM in forecasting the cleaning quality and determining processing parameters, and the mean absolute percent error eMAPE between predictive value and experimental value of water contact angle is 6.09 %. Otherwise, the eMAPE of working current is 4.46 %. Other parameters can also be selected by using this method, and the maximum of eMAPE is 4.458 %. Copyright © 2013 IFSA.



Keywords: Atmospheric pressure, Plasma arc cleaning, Surface quality, Water contact angle.

(ProQuest: ... denotes formulae omitted.) 1. Introduction As a new alternative cleaning method, atmospheric pressure plasma arc (APPA) cleaning can provide a tool to avoid the cited drawbacks of conventional cleaning methods and vacuum limitation of low-pressure plasma cleaning [1-4]. APPA focuses onto the sample surface to generate an energy density enough to induce physical and chemical reactions such as thermal shock, activation decomposition, thermal expansion and spalling [5-8]. The activation decomposition plays a key role in the cleaning process. In addition, the thermal shock of arc energy flow, thermal expansion and spalling of cleaning contaminant can remarkably improve probability of the activation decomposition. As a result, the ablation of the cleaning contaminant takes place at any selected cleaning field. Furthermore, different materials can be cleaned by choosing respective arc energy density without the damage of metal substrate.


However, many factors affect the quality of APPA cleaning, such as the effective power of atmospheric pressure plasma arc, the arc spot diameter, scanning speed, the gas flow rate, the arc current, the distance between the nozzle and the workpiece, the pollutant type and thickness, substrate material properties and geometric dimensions, excitation intensity and excitation waveform of the external transverse alternating magnetic field. In addition, the impact trends of various process parameters on the cleaning quality of APPA, and the interaction of different parameters are complex, which cannot be described by a simple analytic function. Thus to establish a rigorous mathematical model on describing the impact of the process parameters on the cleaning quality of APPA is more difficult.

In recent years, Support Vector Machine [9, 10] has been applied in the field of classification and regression forecasts. It is based on structural risk minimization criteria, and its topology determined by the support vector. Therefore, it can overcome the shortcomings of artificial neural network based on empirical risk minimization criteria [11], and can solve the problem of small samples, nonlinearity, high dimension, etc. The Least Squares Support Vector Machine (LS-SVM) [12, 13] is a new extension of the standard SVM, which can change inequality constraint of SVM to equality constraint. Therefore, quadratic programming problem in the SVM is transformed into solving linear equations. So, LS-SVM can effectively reduce the complexity of solving problems and the computation time.

According to the fact that APPA cleaning being a complicated and non-linear process, cleaning quality being influenced by process parameters and the interaction of different parameters being complex, based on many experiments on APPA cleaning, an intelligent predictive model of the non-linear relationship between cleaning quality (water contact angle on cleaned stainless steel surface) and process parameters is established with the research of Least Squares Support Vector Machines. In this paper, models of surface quality prediction and processing parameter determination on APPA cleaning were constructed based on LS-SVM. For this purpose, a device of APPA cleaning was constructed and applied for the cleaning of 316L stainless steel. The training data was also obtained by using this device. To improve the resulting model's generalization ability, an efficient optimization algorithm known as the crossed grid search method were adopted to tune parameters in LS-SVM design, and a better effect has been obtained.

2. Experimental APPA focuses on the sample surface and generates energy dense enough to induce physical and chemical reactions such as thermal shock, activation decomposition, thermal expansion and spalling. The activation decomposition plays a key role in the cleaning process. In addition, the thermal shock of arc energy flow, thermal expansion and spalling can improve cleaning contaminant capability of the activation decomposition. As a result, an ablation of the contaminant takes place at any selected cleaning field. Furthermore, different materials can be cleaned by choosing proper arc energy density without the damage of metal substrate.

The schematic diagram of APPA cleaning is shown in Fig. 1. The experimental apparatus includes the generator of plasma arc (LHM-30A, Zhongtian Co., China), the coordinate drilling and milling machine (Hanwei Co., China) as the movement of the plasma arc, 316L stainless steel (Huashuo Co., China) as the cleaned workpiece, lubricating oil (Sinopec Co., China) evenly coated on the surface of stainless steel. The APPA cleaning used mixed gas (Ar and 02). Furthermore, the electronic stopwatch (DM3-030, Aoxu Co., China) as timing tool, the electronic balance (FA2004, Hengping Co., China) as weighing tool, which resolution is 0.1 mg and maximum range is 200 g. The water contact angles were measured according to the sessile-drop method using an optical contact angle meter (DSA 100, Kruss, Germany) at ambient temperature. Water droplets (5uL) were carefully dropped onto the surfaces of the plates of 316L stainless steel, and the average value of five measurements obtained at different positions in the samples was adopted as the final contact angle.

According to the principle of APPA cleaning, the process was influenced by more parameters. In this paper, working current I, scanning velocity v, nozzle overhang d, gas rate Q, intensity of the excitation current B were chosen as the most important parameters, as shown in Table 1.

In order to assess the effect of finishing parameters, a five-variable and three-level orthogonal experiment method (LI8(3)5) were conducted. Main effect shows the contribution of individual parameter to water contact angle in the experimental results. All the experimental data of APPA cleaning were list in Table 2, 1-18 tests were used for training data, and 17-20 were used for verifying data.

A very important effect of APPA cleaning is the modification of surface chemical composition. Usually, after APPA cleaning, the surface of substrate is oxidized and become more hydrophilic.

Fig. 2 shows the water contact angle before and after APPA cleaning. It can be seen that the untreated 316L stainless steel surface show a 99.2°water contact angle. After APPA cleaning, the 316L stainless steel surface show a 34.5°water contact angle, indicating that the APPA cleaning can effectively increase the surface energy of the stainless steel. Obviously, the water contact angle in Fig. 2(b) has been greatly decreased compared with which in Fig. 2(a), indicating that the lubricating oil has been cleaned and removed.

Surface contamination can be detected by complementary XPS measurement. The XPS studies were carried out on untreated and treated samples.

Fig. 3 shows the XPS result on the surface of stainless steel including untreated sample (Fig. 3 (a)) and treated sample (Fig. 3 (b)) with APPA cleaning. The XPS spectra of the untreated sample surface showed a high carbon peak due to the high degree of carbon contamination. APPA cleaning of the sample coated surface in an Ar/C>2 discharge remarkably reduces the C peak. Nevertheless, the tremendous decrease in the C signal in the XPS after APPA treatment demonstrates the success of APPA cleaning. However, the oxygen signal of plasma cleaned surface becomes bigger than that of normal stainless steel coated surface. Atmospheric pressure plasma arc cleaning from an Ax/Oi discharge produces oxide growth on the surface.

3. Modeling Approach Principle of LS-SVM: Based on statistical learning theory, VC dimension and structural risk minimization principle, SVM method is a new generation of learning algorithms, with optimization method to solve the problem of machine learning. As a deformation algorithm, for solving SVM, LS-SVM can transform linear inequality constraints of standard SVM into equality constraints and change quadratic programming problems into o linear equations with the introduction of entry of error sum of squares in the objective function. So it can greatly improve the solution efficiency of the SVM, and reduce the learning curve of the SVM.

Consider a given training set D={(xk,>'k), A=1,2,...,N} consisting of N-dimensional input data XkeRN and output data ykeR. The following regression model can be constructed by using nonlinear mapping function (p{x): ...(1) where w is the weight vector and b is the bias term.

By mapping the original input data onto a highdimensional space and constructing a linear regression function, the nonlinear inseparable problem becomes linearly separable in this space. The optimization problem and the equality constraints are defined as ...(2) subject to equality constraints: ...(3) where e^ is the random errors and y is the regularization parameter in optimizing the trade off between minimizing the training errors and minimizing the model complexity.

Then, the objective is to find out the optimal parameters that minimize the predictive error of the regression model, as in Eq. 1. To solve this optimization problem, Lagrange function is constructed as ...(4) where a is the Lagrange multipliers. The solution of Eq. 4 can be obtained by partially differentiating with respect to w, b, e and a.

...(5) then ... (6) With matrix form, the Eq. 5 can be expressed as ...(7) where...is a square matrix, ílu=<p(xk)r<p(x\). I is an identity matrix. Finally, the estimated values of a and b, can be obtained by solving the linear system Eq. 7.

Putting Eq. 6 into Eq. 1, the following result is obtained as ... (8) A kernel function is used as follows: ... (9) Then, the LS-SVM model can be described as follows: ...(10) In the present work, RBF (Radial basis function, RBF) was selected as the following kernel function.

...(11) where o is the kernel width parameter.

Using above LS-SVM model with RBF kernel, the offline nonlinear model of the controlled system can be expressed as follows, and mainly adjustment parameter are y and o.

...(12) In order to acquire a more accurate prediction result, the data set must be conventionally normalized before training [14]. This prevents any parameter from domination to the output value and provides better convergence and accuracy of the learning process. For all input and output values, it is necessary to be normalized within the range [0, 1], through the following transformation formula as ...(13) where «"or is the normalized data set, u is the initial input and output data set, wmin and umax are the minimum and maximum values of the input or output data set.

As all input values are normalized, the output value «"or produced by the LS-SVM is not the actual value. It can be de-transformed using the inverse of «nor'1 in order to obtain the actual output value «.

To achieve a high level of performance with LSSVM model, some parameters have to be tuned, including the regularization parameter y and the kernel parameter cr, as shown in Eq. 12. At present, there is not a unified method for choosing y and o [15]. In this paper, crossed grid search method is applied. The ranges of y and o are selected and the crossed grid is determined. Each crossing point correspond a prediction error. For convenience, o2 is used to be instead of o.

To verify the accuracy of the proposed LS-SVM model, two error functions are established: Mean Absolute Percent Error (cmape) [16].

...(14) where N is the number of validation data, y is experimental value, yy is the LS-SVM prediction value.

4. Results and Discussions Prediction of water contact angle: The input data {xk, yk} of LS-SVM consists of five experimental parameters, including working current (/), scanning velocity (v), gas rate (Q), nozzle overhang (d) and intensity of the excitation current (B). The output data yk is the results of experiments, that is, the water contact angle 0. Thus, the input and output data set are x= {/, v, Q, d, B}, y= {0}, respectively.

The eighteen groups of experimental data selected from Table 2 were used to perform training data, and a training model was constructed.

Toolbox of MATLAB was adopted for the algorithm of LS-SVM, RBF was selected as kernel function, seen in Eq. 11.

A grid search algorithm was applied to obtain optimum value region of the parameter y and o2, y= [0.5, 1, 1.5, ..., 500], oMO.05, 0.1, 0.15, ..., 500]. With the training results of surface quality based on LS-SVM, when the y=4145 and o2=69.32, the least of training error can be obtained.

The rest of four experimental data of Table 2 (19-22) were used for verifying the availability of the prediction model. The processing parameters xk were put into LS-SVM model and the responding result yk is obtained, which was the water contact angle 0. The Comparison of prediction value with experimental value of surface roughness is shown in Fig. 4. The 6mape is 6.09 %.

The actual cleaning process is constrained by the desired surface quality and the treatment cost. Usually, trial-and-error method is used to select the machining parameters. However, it is difficult to obtain the desired surface quality with higher treatment efficiency and a lower cost. Therefore, if a prediction parameter can be selected, a higher surface quality and lower cost can be obtained.

Table 3 shows the factorial analysis of orthogonal experimental of APPA cleaning. Five parameters affect the water contact angle base on priority level: working current I (R=14.850), scanning velocity v (R=9.134), nozzle overhang d (R=7.384), intensity of the excitation current B (R=1.700) and gas rate Q (R=0.0451). It can be seen that the working current has more effect on water contact angle than other processing parameters.

Therefore, working current I was chosen as the output data, water contact angle 0 and other four parameters are regarded as the input data. A new LSSVM training model was constructed to obtain optimum working current which matched with the desired value of water contact angle. With the training results of working current based on LSSVM, when the y=28 and cT=l35, the least of training error can be obtained. Comparison of predictive value with experimental value of working current was shown in Fig. 5. It can be concluded that the eMAPE is 4.46%.

Other processing parameters also can be predicted. Fig. 6 shows the comparison of prediction value with experimental value of scanning velocity, y=215, o2=\3.8, and 6mape is 2.387 %. Fig. 7 shows the comparison of prediction value with experimental value of nozzle overhang, y=56, o2=92, and 6mape is 1.9%. Fig. 8 shows the comparison of prediction value with experimental value of intensity of the excitation current, y=85, o2= 102, and cmape is 1.605 %. Fig. 9 shows the comparison of prediction value with experimental value of gas rate, y=38, o2=268, and ^mape is 2.7 %.

5. Conclusions APPA is applied to clean the lubricant oil covered 316L stainless steel, and a treatment system is also established for obtaining training data. The experimental results show that the atmospheric pressure plasma arc cleaning is effective in reducing considerably the amount of lubricant and increasing the surface energy of the sample significantly in a very short period of treatment time. The water contact angle of treated sample is decreased from 99.2° to 34.5°, and surface energy is improved greatly. XPS results reveal that the surface of stainless steel is oxidized after APPA cleaning, and the mixed gas plasma arc has more ability to reduce the C peak.

LS-SVM prediction model with RBF kernel is applied to predict water contact angle of the sample surface. The results show that, the prediction values based on LS-SVM model are found to be in reasonable agreement with the experimental values. The mean absolute percent error (^mape) is 6.09 %.

The factorial analysis of orthogonal experiments of the APPA cleaning shows that the working current is the most important parameter. Using the LS-SVM model to select the suitable working current and the 6mape is 4.46 %. Other parameters also can be selected by using this method. The maximum of 6mape is 4.458 %.

Acknowledgements The project was supported by the National Nature Science Foundation of China (Grant No. 51205237 and 51375284), also supported by Promotive research fund for excellent young and middle-aged scientists of Shandong Province (Grant No. BS2010ZZ009), respectively.

References [1] . J. B. Meng, W. J. Xu, X. Y. Wang, Modeling of reactive kinetics in the metal surface contaminant cleaning using atmospheric pressure plasma arc, Applied Surface Science, Vol. 254, Issue 21, 2008, pp. 6826-6830.

[2] . S. S. Asada, C. Tenderoa, C. Dublanche-Tixiera, et al., Effect of atmospheric microwave plasma treatment on organic lubricant on a metallic surface, Surface and Coatings Technology, Vol. 203, Issue 13,2009, pp. 1790-1796.

[3] . M. C. Kim, S. H. Yang, J.-H. Boo, et al., Surface treatment of metals using an atmospheric pressure plasma jet and their surface characteristics, Surface and Coatings Technology, Vol. 174-175, 2003, pp.839-844.

[4] . M. C. Kim, D. K. Song, H. S. Shinm, et al., Surface modification for hydrophilic property of stainless steel treated by atmospheric-pressure plasma jet, Surface and Coatings Technology, Vol. 171, 2003, pp. 312-316.

[5] . C. Lee, R. Gopalakrishnan, K. Nyunt, Plasma cleaning of plastic ball grid array package, Microelectronics Reliability, Vol. 39, Issue 1, 1999, pp. 97-105.

[6] . A. Beikind, N. Plainfield, S. Zarrabian, Plasma cleaning of metals: lubricant oil removal, Metal Finishing, Vol. 97, Issue 77, 1996, pp. 19-22.

[7] . L. Canino, G. Napolitano, L. Sorrentino, Correlation of wettability and superficial cleaning of 2024 aluminum alloy with air cold plasma treatment time, The International Journal of Advanced Manufacturing Technology, Vol. 26, Issue 9-10, 2005, pp. 1433-3015.

[8] . J. H. Hsieh, L. H. Fong, S. Yi, Plasma cleaning of copper leadframe with Ar and Ar/Fh gases, Surface and Coatings Technology, Vol. 112, Issue 1, 1999, pp. 245-249.

[9] . C. H. Zheng, L. C. Jiao, Y. Z. Li, Support vector classifier based on principal component analysis, Journal Systems Engineering and Electronics, Vol. 19, Issue 1,2008, pp. 184-190.

[10] . V. N. Vapnik, The nature of statistical learning theory, World Scientific Publishing, Singapore, 2000.

[11] . A. Bapari, A. Najafizadeh, M. Moazeny, Prediction of hot flow stress of CrMoV steel using artificial neural network (ANN), ISIJ International, Vol. 47, Issue 8,2007, pp. 1126-1130.

[12] . D. F. Shi, N. N. Gindy, Tool wear predictive model based on least squares support vector machines, Mechanical Systems and Signal Processing, Vol. 21, 2007, pp. 1799-1814.

[13] . K. W. Pak, V. M. Chi, W. C. Hang, Modeling and prediction of spark-ignition engine power performance using incremental least squares support vector machines, in Proceeding of the 2nd Conference Symposium on Computational Mechanics, Hong Kong, China, 30 November-3 December 2009, pp. 179-184.

[14] . D. Pyle, Data preparation for data mining, Morgan Kaufmann Publishers, USA, 1999.

[15] . M. T. Gencoglu, M. Uyar, Prediction of flashover voltage of insulators using least squares support vector machines, Expert Systems with Applications, Vol. 36, Issue 7,2009, pp. 10789-10798.

[16] . N. S. Patil, P. S. Shelokar, V. K. Jayaraman, Regression models using pattern search assisted least square support vector machines, Chemical Engineering Research and Design, Vol. 83, Issue 8, 2005, pp. 1030-1037.

Xiaojuan DONG, Jianbing MENG, Xiuting WEI, Zhanmin YIN School of Mechanical and Engineering, Shandong University of Technology, Zibo, 255049, China Tel.: 0533-2786909, fax: 0533-2786910 E-mail: [email protected] Received: 16 September 2013 /Accepted: 22 November 2013 /Published: 30 December 2013 (c) 2013 International Frequency Sensor Association

[ Back To TMCnet.com's Homepage ]